Agentic AI Changes Organizational Design, Not Just Workflows

I counted eleven blog posts on agentic AI market sizing and vendor positioning before I gave up. Billions of dollars, percentage gains on customer service costs, which vendor ships agents fastest. Not one of them asked what happens to the people who currently hold the decision authority that those agents are about to absorb. The market sizing framing treats agentic AI as a workflow optimization problem. Automate the task, cut the cost, ship the agent. But workflow optimization is not what is happening. What is happening is a redistribution of who decides, who handles exceptions, and who is accountable when the exceptions nobody planned for surface. That is not a workflow change. That is an organizational design change.

Kumar et al. (2026) lay this out directly. Their paper positions agentic AI as a distinct IS artifact, not an incremental upgrade to automation. Agentic systems, they argue, pursue goals autonomously, coordinate across systems, and adapt over time with minimal human intervention. They distinguish these systems from business process automation, which codifies repeatable tasks within deterministic logic flows and static process maps. Agentic AI shifts from process-centric automation to agency-centric orchestration, where the system itself becomes a decision-making participant. The paper introduces a tradeoff framework built on socio-technical systems theory, and every single tradeoff, from efficiency versus labor displacement to automation versus explainability, is fundamentally about how authority and accountability get redistributed between humans and machines. The workflow is the surface. The authority structure is the depth.

This is where routine dynamics comes in. Feldman and Pentland (2003), as I understand them from the broader IS literature and the way Leonardi (2011) cites them in his study of flexible routines and flexible technologies, gave us the distinction between the ostensive routine and the performative routine. The ostensive is what the organization says it does. The performative is what actually happens when specific people execute the work on a specific day with specific constraints. The gap between them is not a deviation. It is where most organizational value lives, because the exceptions, judgments, and local adaptations that people make are precisely what keep routines from collapsing when reality does not match the process diagram. I wrote about this in the context of process change, but the implication for agentic AI is more unsettling than I initially let on.

When an AI agent takes over a routine, what exactly does it take over? It absorbs the ostensive routine. It handles the standard cases, the documented steps, the predictable flow. It cannot handle the performative variation because the performative is what emerges when people improvise around constraints that were never documented. The agent gets the script. The humans keep the exceptions. And here is the organizational design problem: most middle management roles exist precisely because of those exceptions. Their authority, their expertise, their reason for being in the chain is to handle what the ostensive routine cannot anticipate. If you automate the ostensive routine, you do not just eliminate tasks. You eliminate the structural reason for the role that sits between the routine and the exception.

Orlikowski (1992) gave us the framework to see this before it was visible. Her duality of technology says that technology is both a product of human action, designed and built through organizational processes, and a medium of human action, shaping what users can do. She later called this technology-in-practice, the enacted structure that emerges through recurrent use. Two workgroups using the same system can develop entirely different technologies-in-practice because their patterns of use differ. What happens when one of the actors in that practice is not human? What happens when the agent produces outputs that become inputs to the next cycle of organizational action, and human responses become data that trains or tunes the agent? The recursive loop is still there, just as I argued when I wrote about structuration and AI. But the agency flowing through it is no longer only human.

Giddens (1984) built structuration theory around the idea that structures are both the medium and the outcome of human action. People draw on rules and resources to act, and their patterned action reproduces or transforms those very structures. The three modalities, signification, domination, and legitimation, link structure to agency. Signification produces shared meaning through interpretive schemes. Domination allocates resources to produce power relations. Legitimation deploys norms and sanctions to produce moral order. When an AI agent starts making scheduling decisions, filtering support tickets, approving low-risk transactions, and flagging which projects get funded, it is not just performing tasks. It is producing interpretive schemes, because its categories and classifications become how the organization thinks about priority. It is allocating resources, because its decisions determine who gets attention and what gets done first. It is deploying norms, because its validation signals and rejection patterns reinforce what counts as acceptable. As I wrote about same tool, different outcome, the three modalities are where the real action is. The agent is not a neutral tool. It is a structural participant.

DeSanctis and Poole (1994) add the concept of spirit and appropriation. Technologies come with a spirit, the designer's intent for how they should be used, and structural features, the capabilities embedded in the system. Groups appropriate those structures, and the appropriation can be faithful or unfaithful. The question for agentic AI is who does the appropriating when the system is actively generating outputs that shape behavior. When a scheduling agent optimizes for back-to-back efficiency, its spirit is embedded in its optimization criteria. The organization may have a completely different spirit for what meetings should look like, but the agent's outputs, its classifications, its priority rankings, its scheduling logic, become part of the structure before anyone has a chance to appropriate faithfully or unfaithfully. The appropriation happens at the pace of algorithmic output, not at the pace of organizational reflection. I think this is the most undertheorized aspect of agentic AI adoption right now. The delegation problem that Baird and Maruping (2021) identified is not just about measuring use differently. It is about who holds the appraisal, distribution, and coordination functions in the organization. Before agentic AI, a manager appraised a task, decided who should do it, distributed it, and coordinated the interdependencies. When the agent takes over distribution and coordination, the manager's role shrinks to handling exceptions. The structural position of middle management shifts from decision authority to exception processing. That is not a workflow change. That is a redefinition of what the role exists to do.

Leonardi (2025) names this directly. He argues that when humans attribute agency to AI systems, they experience power displacement, meaning a reduced sense of control and responsibility even while retaining formal authority. The human still has the title. The human is still accountable on paper. But the felt experience of agency, the sense that you are the one causing outcomes, has shifted to the machine. His agency loops concept describes how this displacement circulates. A manager delegates to the agent. The agent acts. The manager attributes the outcome to the agent. The manager then adjusts future behavior based on that attribution. The loop reinforces itself. The manager retains formal authority but loses the felt experience of being the one who decides. Over time, the organizational habit of deferring to the agent becomes structural. Nobody made a decision to restructure. The routine changed, and the structure followed.

I think the reason nobody writes about this is that organizational design sounds slow and boring compared to market sizing. A $376 billion market projection makes for a better blog post than an argument that middle management roles get hollowed out when agents absorb the routine decisions and leave only the exceptions, which are the decisions that require the most judgment, carry the most risk, and come with the least precedent. The fun part about automating the routine is that you get efficiency metrics on a dashboard. The uncomfortable part is that you have also removed the training ground where people learned how to make decisions in the first place. Routine decisions are how junior managers build judgment. They are the ostensive practice that develops performative skill. Automate those, and you do not just eliminate tasks. You eliminate the developmental path that produces the people who can handle the exceptions.

Kumar et al. (2026) flag this in their tradeoff framework. Their efficiency versus labor displacement tradeoff notes that agentic AI does not simply replace jobs but reallocates tasks across humans and machines, and that the new tasks generated by agentic systems, oversight, governance, exception handling, are unevenly distributed and require different skills. The people who lose routine tasks to agents may not be the same people who gain governance tasks from agent deployment. The organizational chart does not just shrink. It reshapes. And routine dynamics tells us that the reshaping will not be planned. It will emerge from the gap between what the agent is designed to do and what the organization actually needs, which is always more than what fits in a process diagram.

The practical question is not whether agents should be deployed. They will be. The question is whether the organizations deploying them understand that they are not just automating workflows. They are restructuring decision authority. They are changing who gets to decide what, who handles the exceptions, who is accountable when the agent makes a decision that the process diagram did not anticipate. Structuration theory predicts that these changes will compound. The agent's outputs become part of the structure that conditions future action. The habitual deferral to the agent becomes a norm. The norm becomes a legitimacy structure. At that point, reversing the decision to deploy the agent means reversing a whole organizational routine, not just turning off a tool.